Experimental results under 1x(12ep) training strategy? about queryinst HOT 2 CLOSED

Hi, @Jason-cpp, thanks for your interest in our work!

1x schedule is originally designed for Faster / Mask R-CNN families, optimized using SGD with a learning rate 0.01.

While Transformer / Query based e2e object / instance level frameworks like DETR, Def-DETR & Sparse R-CNN are optimized via Adam families using a very small learning rate. Usually, it will take much longer for them to converge.

To my knowledge, under 1x schedule, the ranking of Box AP is: Faster / Mask R-CNN >= Sparse R-CNN >= Def-DETR >> DETR. If you are interested in the precise numbers, I suggest you open an issue in the corresponding repo.

Since the 1x schedule cannot provide good enough detection performance, it is not so much meaningful to consider 1x instance segmentation performance for a proposal based method.

Also, QueryInst is designed to be performance-oriented, i.e., we are not so much interested in the 1x performance, but we do care about the fully converged model performance.

We argue that 1x is not a "standard schedule" for every model and it is too harsh to compare every model under the 1x schedule (1x is too short !). A good 1x model is not guaranteed to be a good (fully converged) model, e.g., EfficientDets & YOLOS families all need hundreds of epochs to fully converge.

from queryinst.

We believe we have answered your question, and as such I'm closing this issue, but let us know if you have further questions.

from queryinst.